Power budget for a home server: PSU headroom and UPS decisions

The power budget is the least glamorous decision in a home server build and the one most people get backwards on the first attempt. The usual sequence runs like this: add up nameplate ratings to pick a power supply, buy a UPS because the box says it will run a whole PC, then discover six months in that the server idles at a fraction of what you sized for while the battery degrades faster than expected in the closet. None of that is the hardware’s fault. The two numbers that actually govern both decisions, how much the machine draws at the wall and how long the backup needs to run, are exactly the two numbers that most builders never measure.

This guide walks through the power decisions in the order they should be made: how to measure what your build really draws, how PSU headroom is supposed to work instead of being a vague cushion, why UPS sizing runs on watts rather than volt amperes, which loads you actually size for, how to read a runtime curve against a shutdown window, and what the battery replacement cycle costs so it lands in the budget as a line item instead of a surprise. By the end you will be able to write down a defensible power budget for any future home server before you buy a single component.

Start With What the Server Actually Draws

Every power decision downstream depends on one measurement: what the build draws at the wall under real conditions. A plug in energy meter or a smart plug left on the server’s input for a week gives you the honest figures, idle draw and peak draw, and the numbers will be far smaller than the sticker on the power supply. A server whose 650 watt PSU nameplate suggests a hungry machine typically idles at 40 to 60 watts and peaks around 150 to 250 watts, because the nameplate reflects worst case rail capacity, not what the hardware consumes running services. A decent build of one server, a NAS, a switch, and a router settles into steady state somewhere between 100 and 180 watts at the wall.

The measurement has two wrinkles worth remembering. Disks are the noisy variable: every spinning drive adds 5 to 9 watts in steady state and spikes to roughly 25 to 30 watts each for a few seconds at power on. And the wall reading already includes PSU conversion losses, because a 90 percent efficient supply draws about 11 percent more than the components actually use, so what you read from the meter is exactly the number the UPS must carry. Write down both the idle figure and the peak figure, because the idle figure drives runtime planning and the peak figure drives sizing, and they are not the same number.

PSU Headroom Is an Efficiency Question, Not a Cushion

A power supply converts wall AC into the DC rails a server needs, and every conversion loses a little energy as heat. The loss is not constant: efficiency climbs from low load toward a peak at roughly 50 percent and falls off again near the top, which is why a Gold rated supply hits about 90 percent around half load and drops back to 87 percent at both 20 and 100 percent. Titanium, Platinum, and the other tiers shift that whole curve up a few points, and for a home server the tier difference is worth a few dollars a year at most, not worth rerunning the build. What matters is where your load lands on the curve, not which badge the supply carries.

That reframes the headroom question. You need headroom so transient loads, drive spinup, CPU boost, a future GPU, never push the supply past its rating and trip its overcurrent protection, but you can also buy too much headroom. A 1000 watt supply powering a 40 watt server sits near 4 percent load, at the worst point on the efficiency curve, and it costs extra money to raise your idle losses. A practical target is peak hardware draw of 60 to 70 percent of the PSU rating with typical draw somewhere above 15 to 20 percent, which for the usual 150 to 250 watt peak home server means a 300 to 450 watt supply, and a 500 to 650 watt unit only if a GPU or a large disk shelf is genuinely in the plan.

Size the UPS in Watts, Not VA

The UPS box carries two numbers, and the bigger one is the one you should ignore. Volt amperes describe apparent power, the voltage times the current the unit can deliver, while watts describe the real power it can sustain, and the two are linked by a power factor that is usually below 1.0 on budget towers, which is why a “1500 VA” unit is frequently rated at only 900 watts. Your server’s own active power factor correction supply draws close to its real wattage, so your load is watts and the UPS’s watt rating is the binding constraint. Sizing to VA means trusting a number that neither your equipment nor the UPS meaningfully uses.

The sizing rule is short: take the total measured wall draw under realistic active load, add 20 to 30 percent headroom for spikes and growth, and choose a UPS whose watt rating comfortably exceeds that figure. The load should sit below 60 percent of the rating whenever possible, which is where runtime behaves predictably, and it should never wander near 90 percent because runtime collapses at the top of the curve and the overload buzzer starts competing with the server for your attention. For the typical home server stack that means the 900 watt class of tower UPS, the ubiquitous 1500 VA units, is the sweet spot for most builds, stepping to 1200 watts only for a GPU node or a second host.

Plan for Steady Draw, Peaks, and Spinup

A power budget rests on three different loads, and conflating them is where most sizing errors come from. Steady draw is what the build pulls the vast majority of the time, and it is the figure you multiply by to estimate daily energy and the figure the UPS runtime curve is read against. Peak draw is the sustained maximum, usually reached during a backup job, a transcode, or a rebuild, and it is the figure the UPS watt rating must cover with headroom. Spinup is the transient: when power returns after an outage and the server boots from battery, every drive spins up at once and the draw briefly jumps above steady state before settling back down.

Each load has its own consequence. If your peak draw pushes past the UPS watt rating, you get overload beeps and the UPS transfers to battery at the worst possible moment, so size the watts to the peak, not the idle. If spinup can happen on battery, the UPS must ride that spike, which is a real argument for staggering drive spinup where the controller supports it and for keeping the boot draw inside the watt rating with room to spare. And a simple rule protects the path: nothing with a motor belongs on battery outlets, so printers, heaters, and similar loads stay off the UPS entirely, leaving battery capacity for the equipment that actually needs a clean shutdown.

Read the Runtime Curve Against the Shutdown Window

The runtime figure printed on the box is the misleading one, because vendors quote it at full rated load, a state your build essentially never reaches. The reality is a steep nonlinear curve: the same 900 watt tower that runs about 3 minutes at full load stretches to 8 minutes at half load and 35 minutes or more at a typical 180 watt idle. Halving the load does not halve runtime, it more than doubles it, which is why an apparently small UPS at a small real load can still feel like a very long runway. Every model publishes its own curve, and the only runtime number that matters is the one read at your actual draw.

The right sizing target is not maximum minutes, it is the shutdown window. A clean shutdown is a software activity: on Linux, Network UPS Tools, on many NASes their built in integration, detects the outage, waits through a configured grace period in case mains returns quickly, then tells the services to stop cleanly before the battery dies. That sequence typically needs 5 to 10 minutes from the moment the outage is confirmed, so a runtime of 15 to 20 minutes at your idle draw covers the window with margin and keeps the battery from being deep cycled. A UPS without a monitoring link is a very expensive power strip, because nothing will ever tell the server to shut down on its own.

Budget for the Battery Replacement

The battery is a consumable, and pretending otherwise is how a UPS becomes a surprise. Sealed lead acid batteries inside consumer towers last roughly three to five years, and the countdown runs faster in any closet warmer than room temperature, which most server closets are. The failure mode does not announce itself: runtime simply shrinks over time until one outage exceeds what the battery can carry. The cheap insurance is a runtime test at the start of year three, compare the result against the vendor curve at the same load, and replace the cartridge once runtime drops below about 60 percent of the rating when new, or at year four or five, whichever comes first. The cartridges themselves are commodity parts that swap in minutes and cost tens of dollars, which works out to a few dollars a month of amortized budget.

The battery cost is also the hidden argument against oversizing. A bigger UPS than you need means a bigger and pricier battery cartridge to replace every few years, sitting lightly loaded for a near constant cost, so the right sized unit wins on battery economics as well as on the runtime curve. Lithium designs such as LiFePO4 last longer and degrade more gracefully, but they cost several times more up front and remain rare in consumer UPS towers, so for a home server the lead acid unit with a monitoring link and a replacement plan is still the sensible default. Whatever you choose, write the replacement date and the test cadence down, because the battery that fails in year four was predictable in year three.

Turn the Budget Into a Record

A power budget is only as trustworthy as the measurements feeding it, and the measurements only survive if they live with the hardware. Record the measured draw on each device record next to the serial number, model, and role, note which UPS feeds which equipment, and write down the tested runtime from the day the system was commissioned, because that number is your baseline for the year three battery check. The same habit keeps the budget honest when the build changes: adding a drive, swapping a host, or enabling a new service moves the draw, and a record that still says what the old hardware drew will quietly invalidate every future decision.

A home server budget kept this way is a small model of the larger discipline that rack power budgeting applies to cabinets and circuits, where measured draw, headroom, and redundancy are tracked against a documented baseline.